geostatistical tools for analyzing spatial extreme...

1

Geostatistical tools for analyzing spatial extreme rainfall

patterns over Tunis City

Zoubeida BargaouiNational School of Engineers of Tunis (ENIT)

Tunis El Manar University

4th International workshop on hydrological extremes Med-FRIEND group

University of Calabria 15-17 September 2011

Zoubeida.bargaoui@laposte.net

2

Introduction• The identification of the structures of heavy rainfall

spatial variability helps:– deriving rainfall patterns – quantifying rainfall amounts in unobserved sites – Areal rainfall estimations are quite important for flood and

erosion risk assessment as well for economic damage evaluation

• Geostatistical tools are worth noting to address these concerns through

– Assessment of variogram functions to identify the variability structure of a given rainfall pattern (as a random field).

– And to perform Kriging as optimal interpolation method

3

Variogram assessment• Commonly, geostatistical approach relies on the

identification of the variogram function γ(h) which represents the mean quadratic deviation in the random field

• γ(h) = ½ E[(Z(x)-Z(x+h))²]

• Sample variogram γexp(h):• γexp(h) = ½ ∑[(Z(x)-Z(x+h))²]/N(h)• N(h) number of pairs for a given class of distance with

average value h.

• Robust estimator (Cressie, 1993) adopts the absolute deviation

• γrob(h) = ½ (1/(0.457+0.494/N(h))∑[| (Z(x)-Z(x+h)) |]

4

Drawbacks of variogram function• However the assessment of rainfall spatial dependence

attached to heavy rainfall events needs specific approaches because variogram analysis assumes that the marginal distribution of the underlying random field is Gaussian

• Rivoirard (1991) proposed disjunctive kriging which corresponds to a cokriging of indicator function as a method of non linear geostatistic.

• This method was addressed as alternative to kriging of logarithmic transformation of data because of problems rising from the back transformation to original data

• Rivoirard J. (1991) Introduction au krigeage disjonctif et à la géostatistique non linéaire. Cours Centre de géologie appliquée. Ecole des Mines de Paris.

5

• The gaussian type of dependence underestimates the dependence of extremes

• Cooley (2005) suggested that «the madogram ν(h)conveniently links geostatistical ideas to measures of dependence for extremes » ν(h)= ½∑ |(Z(x)-Z(x+h))| / N(h)

• and that «the madogram could be used to model the spatial dependence as a function of distance, much like the variogram»

• The madogram was proposed by Matheron (1987) in linkage with the variogram of the indicator variable

6

• KazienKa and Pilz (2010) mentioned well-known methods to deal with non- gaussianity (kriging of logarithmic transformation of data ; disjunctive kriging; rank-order transformation)

– Kazienka H.., J. Pilz., (2010), Copula based geostatistical modeling of continuous and discrete data including covariates. Stoch Environ Risk Assess 24: 661-673.

• Smith (1990) proposed multivariable t-transformation as extreme value process

• Copula function has been proposed as alternative and Bardossy and Li (2008) proposed a copula function based on Khi-Deux transformation.

– Bàrdossy A, Li J (2008) Geostatistical interpolation using copulas. Water Resour Res 44:W07412

• They demonstrated that Indicator kriging ; disjunctive kriging; simple kriging of the rank-transformed data are Copula based spatial interpolation models

7

joint distribution (bivariate) and copula

• Copula is defined as a distribution function with marginal distributions uniforms on (0,1).

• Sklar theorem (1959) links copulas and multivariate distributions F with margines F1 and F2.

• C(u1, u2)=prob (F1(X)≤u1, F2(Y) ≤ u2)• If the distribution is continue then the

copula is unique

8

Aim of this work

Achieve a quantitative analysis of tail distribution of heavy rainfall eventsusing geostatistical tools

– and Generalized extreme value (GEV) bivariate distribution

• Case study: Analysis of Tunis daily precipitations patterns. September 17th 2003

9

Data

• maximum annual daily precipitation recorded at 12 stations sample sizes (16 to 116 years).

• precipitation records on 17/9/2003 at 75 stations max = 187 mm/day

10

FIG 1 Fig. 1 Réseau et totaux du 17-9-2003 (mm)

4000000

4020000

4040000

4060000

4080000

4100000

4120000

4140000

550000 560000 570000 580000 590000 600000 610000 620000 630000 640000 650000 660000localisation longitude X(m)

Loca

lisat

ion

latit

ude

Y(m

) Série1> 100 mm80-100 mm50-80 mm

pas de pluie

pas de pluie

pas de pluie

Ariana

Tunis Carthage

Fondouk Jedid

La Manouba

Gammarth

Uthique

Zaghouan

Mer Mediterranée

Korbous

DDC

Tmanou

11

FIG 2series of concomitant data (excluding gaps) for Tunis Manoubia and Borj Chakir stations

0

20

40

60

80

100

120

140

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91

years

daily

max

ann

ual r

ainf

all (

mm

)

Borj ChakManoub

12

A) Application of Extreme value theory

• C(u,v)=prob (F(X)≤u, G(Y) ≤v)= exp [log(uv) A (log(u)/log(uv))]

• A is the Pickands dependence function• (Gudendorf and Segers, 2009)

• This function has been extended in the bivariate case (Cooley;2005)

• Z(x) et Z(x+h) • C(u,v,h)=prob (F(Z(x))≤u, F(Z(x+h)) ≤v)= exp

[log(uv) A (log(u)/log(uv), h)]

13

Definition of extremal function θ

θ(h) is defined as the tail of the joint bivariate distribution :

• Prob(F(Z(h))≤u, F(Z(x+h)) ≤u) = uθ(h)

θ(h) gives a partial idea on the tail dependence structure (conversely to A);

• Extreme dependence for θ(h) =1 ; • Extreme independence for θ(h) = 2

14

Example of θ estimation• Smith (1990) assumed local GEV model (µ,σ, ξ)

for the variable X, • he adopted a standardisation of data through the

Frechet transformation y = (1 −ξ(x − µ)/σ))−1/ξ

• With Pr{Yt ≤y} = exp(−1/y)• In this case, 1/Y has unit exponential

distributions and • 1/ max(y1,y2) has an exponential distributionwith mean 1/θ• He plotted the extremal function as function of h.

15

Extremal dependence function and Pickands function

• In terms of the dependence function A, the extremal dependence function is (De Haan,1985) θ(h)= 2 A(1/2,h) (1≤ θ(h) ≤2)

• Etimation of A(t,h) and A(1/2,h) to derive θ(h)

16

Non parametric estimation of Hall and Tajvidi (2000) for A

• recommanded by Gudendorf and Segers (2009)

ς HTi(t)= min [-log(u)/LUM/ (1-t), -log(v)/ LVM/ t ] ; 0≤ t≤ 1 – LUM average value of – log(u) – LVM average value of – log(v)

• Where u and v are Frechet transforms• k• 1/AHTest(t) =1/ k ∑ ς HTi(t)• i=1

17

B) Analysis of scales of variability according to the results of the calibration of variograms and

madograms• Use of Simulated annealing (Monte Carlo simulation

optimization method) with RMSE as calibration criteria.• Changing the seed to initiate SA (replications) and

multiplying the SA calibrations, it is possible to investigate the uncertainty about variogram (madogram) model.

• SA: annealing temperature decrease is according to a geometric distribution Tn=T0 (RT)n

• Stopping criteria Tn< Tratio with T0 =10-6

• RT is computed according to the maximum number of itérations ITmax with no change in temperature

• T0=0.2 ; ITmax = 70 so that RT=0.7.

18

C) Derivation of a regional GEV distribution assuming relationship between θ(h) and ν(h)

• For the GEV distribution, θ(h) is related to the madogram (Cooley, 2005)

• So, θ(h) is modeled with a geostatistical spherical model.

• GEV ξ≠0 :θ (h)=[1+ ξ ν(h)/ (σ Γ ( 1− ξ))] 1/ ξ

Γ Digamma function. • Gumbel

θ( h) = exp (ν(h)/ σ)

19

Results• Scatterplots (Fig. 2)• Estimation of the dependence function (Fig. 3)• Estimation and modeling of the extremale function (Fig.

4)• Identification of the variogram and madogram of the

rainfall field (Fig. 5a)• Estimation of marginal GEV• Identification of the variogram of GEV parameters

separately (Fig. 5b,c)• Comparison of scales of variability• Mapping the conditional joint distribution for a given risk

value (u=0.98) for a given location (Fig. 6)

20

Practical aspects

• 1 km grid size (Number of nodes = 7793). • inter distance for sampling variogram and

madogram: 3.5 km. • 18 classes of distances.

21

Scatterplots of data and Transformed data (FIG 2)

Transformed v (Potin B.) versus transformed u (T. Manoubia) θ=1.44

0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1

197

(u,v) for the pair Tunis Carthage and Tunis Manoubia θ=1.56

0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1

u Tunis Manoubia

v Tu

nis

Car

thag

e

2003(1)1975(2)

1979(4)1973(3) 1976(5)

1969(13)

T Carthage vs T Manoubia data r=0.51

0

20

40

60

80

100

120

140

160

180

200

0 20 40 60 80 100 120 140

T. Manoubia (mm)

T. C

arth

age

(mm

)

T. Manoubia vs Potin B. r=0.53

0

20

40

60

80

100

120

140

0 20 40 60 80 100 120 140

Potin B. (mm)

T. M

anou

bia

(mm

)

197

22

FIG 2 (next)T. Manoubia vs Borj Chakir r=0.67

0

20

40

60

80

100

120

140

0 20 40 60 80 100 120 140

Borj Chakir (mm)

Tuni

s M

anou

bia

(mm

)

Scatterplot of u,v for Borj Chakir and Tunis Manoubia stations

0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1

0 0,2 0,4 0,6 0,8 1u BC

v TM

θ=1.38

23

Fig. 3

Fig. 3. Fonction de Pickands pour les paires de stations

0,65

0,7

0,75

0,8

0,85

0,9

0,95

1

0 0,2 0,4 0,6 0,8 1t

A(t)

Mnih-TCAria-TCDD-TCTM- DDDD-ChTM-TCTM-POTM-BCTC-ChTC-POTM-MnihPolynomial (Mnih-TC)Polynomial (TM-TC)Polynomial (Mnih-TC)Polynomial (TM- DD)Polynomial (TM-BC)Polynomial (Aria-TC)Polynomial (DD-TC)Polynomial (DD-Ch)Polynomial (TM-PO)Polynomial (TC-Ch)Polynomial (TC-PO)Polynomial (TM-Mnih)

24

Fig. 4 (fitting of a spherical model; 10 km for extreme independency)

Sample and adjusted Extremal function (range parameter =15 km)

1,000

1,100

1,200

1,300

1,400

1,500

1,600

1,700

1,800

1,900

2,000

0 5000 10000 15000 20000 25000 30000 35000 40000 45000Inter distance (m)

θ(h )

93

9

4712

1239

25

36

43

32 30

θ(h)=1.25+0.3*(1,5*(h/15000)-0,5*(h/15000)3)

25

Partial conclusions on scales of variability according to 10 replicates of SA

• Classic and robust variograms lead to different results for the rainfall fied and the scale parameter σ field but there is insensitivity for ξ.

• Scales of variability (according to the decorrelation distance)

• Range parameter for ξ and rainfall field are of the same order (optimal values: 17 to 20 km; fluctuation of accepted solutions between 12 and 24 km).

• Conversely, for σ optimal range is 40 km and fluctuation is larger : 30 to 99 km).

• No study of µ yet

26

FIG 5

• .

Fig. 2 Estimation du variogramme de la pluie

0

500

1000

1500

2000

2500

3000

3500

0 10000 20000 30000 40000 50000 60000 70000

inter distance (m)

mm

²GG RobusteSIM1 RSIM2 RSIM3 RSIM4 RSIM5 RSIM6 RSIM7 RSIM8 RSIM9 RSIM10 RSIM1 SIM2 SIM3 SIM4 SIM5 SIM6 SIM7 SIM8 SIM9 SIM10

Fig. 3b Variogramme de Sigma

0

10

20

30

40

50

60

70

0 10000 20000 30000 40000 50000 60000 70000

interdistance (m)

Sample variogramSIM1 RSIM2 RSIM3 RSIM4 RSIM5 RSIM6 RSIM7 RSIM8 RSIM9 RSIM10 R

Fig. 3 Variogrammes et ajustements sphériques

0

0,005

0,01

0,015

0,02

0,025

0,03

0,035

0,04

0,045

0,05

0 10000 20000 30000 40000 50000 60000 70000

Interdistance (m)

γ(ξ)

SampleSample RobustSim1_RobSim2_RobSim3_RobSim4_RobSim5_RobSim6_RobSim7_RobSim8_RobSim9_RobSim10_RobSim1Sim2Sim3Sim4Sim5Sim6Sim7Sim8Sim9Sim10

Fig. 3c Estimation du madogramme

0

5

10

15

20

25

30

35

0 2 4 6 8 10 12 14 16 18 20

Interdistance

Mad

ogra

mm

e (m

m²)

27

Scales of variability according to atmosphere dynamics

Variability of ξ and extreme 2003 rainfall field: micro scale; local effectsVariability of σ: mesoscale; dictated by General atmospheric circulationVariability of θ (10 km) microscale effect

28

FIG. 6 (using the θ spherical model)The tail of the joint bivariate GEV distribution Prob(F(Z(h))≤u, G(Z(x+h))

≤u) = uq(h); Z(h):Tunis Carthage (Airport) station; u=0.98

4000000

4020000

4040000

4060000

4080000

4100000

4120000

4140000

550000 560000 570000 580000 590000 600000 610000 620000 630000 640000 650000 660000

independence0.03590 <P < 0.035950.03575 <P < 0.035900.0350< P < 0.035750.0335 < P < 0.03500.0325< P < 0.03350.0300 < P < 0.03250.0285< P < 0.03000.0250< P < 0.02850.020 <P < 0.0250Stations with seriesObservation sites

29

Conclusions

• raw estimates of the extremal coefficients (not based on any model) were derived

• Estimation of standard error of θ may be achieved through Jackknife procedure (by omitting observations of a year, year by year)

• Attention is restricted to extremes based on annual maxima but the development of threshold methods should be investigated